Dimitris Papailiopoulos
@DimitrisPapail
Followers
20K
Following
12K
Media
947
Statuses
9K
Researcher @MSFTResearch, AI Frontiers Lab; Prof @UWMadison (on leave); learning in context; thinking about reasoning; babas of Inez Lily.
Madison, WI
Joined May 2012
Thinking Less at test-time requires Sampling More at training-time! GFPO is a new, cool, and simple Policy Opt algorithm is coming to your RL Gym tonite, led by @VaishShrivas and our MSR group: Group Filtered PO (GFPO) trades off training-time with test-time compute, in order
19
43
364
I am afforded to make the speculation because I had the best advisor.
0
0
9
as with many things randomness plays a huge factor. And the highest predictive feature of success is not who your advisor is, but how strong of an intellectual connection and mentoring relationship you can develop with them.
1
2
19
A PhD can be incredibly rewarding and fun, under the right circumstances. If you have tremendous *fire* for research and self motivation, it can be the right path for you. If you're doing it because it's the least friction next step after ugrad or MS, reconsider.
I still get a lot of questions about doing a PhD in ML so I'm resurfacing a three threads I wrote about this including some fun personal stories š§µ
8
27
405
mediocre human writing > polished llm writing
Skimmed through a few ICLR26 submissions on OpenReview. Way too many titles are so obviously LLM generated. Advice for junior authors: do not use LLMs for titles, abstracts, or anything related to your story telling. Your writing voice matters!!! It signals taste and thinking.
3
1
41
Skimmed through a few ICLR26 submissions on OpenReview. Way too many titles are so obviously LLM generated. Advice for junior authors: do not use LLMs for titles, abstracts, or anything related to your story telling. Your writing voice matters!!! It signals taste and thinking.
12
23
324
Say a startup raises O(1B). They do a lot of creative and cool things, but they burn the money without delivering monetizable products. What happens next? Do they just apologize to the investors and move on?
8
0
68
āā¦no matter how much math we use, generalization theory is not predictive of what happens in practiceā
Almost a decade ago, I coauthored a paper asking us to rethink our theory of generalization in machine learning. Today, Iām fine putting the theory back on the shelf.
6
4
113
looped transformers strike back!
New paper š: Tiny Recursion Model (TRM) is a recursive reasoning approach with a tiny 7M parameters neural network that obtains 45% on ARC-AGI-1 and 8% on ARC-AGI-2, beating most LLMs. Blog: https://t.co/w5ZDsHDDPE Code: https://t.co/7UgKuD9Yll Paper:
0
6
57
@tszzl @j_mcgraph @SebastienBubeck i beg you plz share with someone. the app rendering is driving me nuts, and the web version is insanely slow when context is very long :D neither works!
2
0
14
Hey, New York ā Brian thinks he's our new CFO. We gave him a stage to prove it.
10
4
18
Can someone at OpenAI please fix the LaTeX rendering in the chatgpt app? The website renders fine, but the app is broken. PLEASE FIX
10
3
67
Advisor models seem v. cool to me in terms of guiding larger, closed models. The advisor seems like a prompt optimizer, but can be much smaller/cheaper to work with than optimizing prompts directly on the big closed model. RL on the advisor makes a ton of sense too. Cool work!
Training our advisors was too hard, so we tried to train black-box models like GPT-5 instead. Check out our work: Advisor Models, a training framework that adapts frontier models behind an API to your specific environment, users, or tasks using a smaller, advisor model (1/n)!
0
2
23
Super excited to present our new work on hybrid architecture modelsāgetting the best of Transformers and SSMs like Mambaāat #COLM2025! Come chat with @nick11roberts at poster session 2 on Tuesday. Thread below! (1)
2
24
68
MSR NYC is hiring spring and summer interns in AI/ML/RL!
8
27
407
Check Thinky's Tinker codebase. GRPO is out REINFORCE with Adv = Reward-mean(Reward) is in NO CLIPPING model ā model + Ī· Ā· advantage Ā· ā logprob
13
19
457
in NYC a single pour over costs as much as one month of chatgpt plus. Just pause and think about it for a second.
7
0
20
verifiable math is an area were self improvement can occur successfully. Very curious to see where this goes, along with several other concurrent efforts.
Today, I am launching @axiommathai At Axiom, we are building a self-improving superintelligent reasoner, starting with an AI mathematician.
1
2
23
Iāve been hearing that stochastic gradient descent is in fact far behind frontier labs. Is that true??
8
6
327
Prediction: In ~3 years academia will be the most desirable place to do fundamental AI research Contributing factors: - small models improve/become significantly more impactful - open weights community broadens its reach - gpus continue to get faster & cheaper - meaningful
23
34
472
If you stare for a long at the lambda $/GPU/hour charges, you start appreciating the actual value of the chatGPT pro subscription..
5
0
27